3 dumb things people build with AI (and 4 smart ones)

An AI agent misread a reply asking for 4 Solana tokens and sent $441,000 in crypto instead.

It happened this year.

You’ve seen it.

The post by some crackpot who thinks he reinvented the universe by vibecoding a basically direct replica of already existing SaaS.

Meanwhile you’re shaking your head like, what an idiot, I bet it has 0 revenue.

That gap between “can you build it” and “should you build it” is getting expensive fast.

AI made building things nearly free (theoretically), so people are now building things no one asked for, wire them to a live bank account or real customer data…

…and find out later that "it compiled" and "it's safe to run unattended" are two very different claims.

Before you build it, you gotta run it through the filter:

Is it worth it to build this?

Funny enough, it’s the same filter SaaS founders were forced to consider not too long ago.

Why almost nobody runs that filter first

• They see a slick demo and skip straight to production, on money or data that actually matters.
• They confuse "I got it working once" with "this is safe to run unsupervised."
• They build the tool they'd personally want, not the one that other businesses would need.
• They skip the boring parts (fees, spreads, access controls, edge cases) because AI made the fun part so easy.

Once you can spot the difference between a toy and a tool, sorting your own AI wishlist into keep and kill takes about 10 minutes.

The dumbest stuff I’ve seen

If you’ve found better, tag me / message me it on LI.

AI trading algorithms

Building your own AI trading algorithm is one of the fastest ways to lose real money. People do it anyway, because a backtest makes it easy to fool yourself.

Why it's stupid:

• The backtest runs on a clean slice of historical data, prints gains, and builds false confidence before real capital goes in.
• It skips trading fees and bid-ask spreads, so the "proven" strategy is wrong from day one.
• It's overfitting: a strategy tuned to one window of history works for exactly that window, then breaks the moment the market shifts.
• Algorithmic trading systems like this literally do not work - and haven’t worked - in the absence of other extremely specific advantages (like fiber optic lines directly connected to the exchange)

In 2026, Bloomberg tracked AI bots auditioning for real Wall Street trading contests, and most of them lost money.

I've never met a business owner who beat the market with a bot they built themselves, but I've met a few who lost 6 and 7 figures trying. And I’ve watched much-more-sophisticated hedge funds go up in flames thinking their model is smarter than everyone else.

I’d go so far as to say that about 99% of people building trading algorithms are scamming people looking to get rich quick.

Rebuilding existing SaaS products

"I'll just build my own version" sounds cheap when a chatbot writes the code in an afternoon. But the moment you need to scale, all hell breaks loose.

Why it's stupid:

• You end up with a janky product and a pile of tech debt, and there's no paid team on call to fix it when it breaks.
• AI-generated code skips security fundamentals an experienced developer adds without thinking twice: access checks, rate limits, row-level security.
• Best of luck getting integrations to real rails. Your vibe code isn’t going to pass a SOC2 audit.
• None of this shows up in the demo. It shows up 6 months later, during a security review, a breach notice, or a 2am incident.

The tool writing your code doesn't know your business is going to have real users and real data next month. It's built to pass your test, but almost nobody separately checks whether it survives the internet.

It goes so far beyond “hey make sure you don’t expose your keys in your repo.”

If you wouldn't trust yourself to configure a firewall by hand, don't trust an AI agent to do it silently in the background while you're not watching.

Replit's AI coding agent deleted a live production database mid code-freeze, after being told explicitly not to touch it. "This was a catastrophic failure on my part," it said afterward. "I destroyed months of work in seconds."

Oopsie.

Then it told the founder the data couldn't be recovered. He recovered it manually.

LinkedIn and Twitter posting (and commenting)

Half of LinkedIn now reads like it was written by the same 4 prompts. The serious version of this build is the auto-poster, and its little brother, the auto-commenter. You already know its work:

"Great insights. This is exactly why consistency compounds. Thanks for sharing."

You're not fooling anyone.

Why it's stupid:

• Everyone can spot it now: generic phrasing, no real opinion, comments that could sit under any post.
• LinkedIn is rolling out a detector that quietly caps how far flagged posts spread outside your own connections.
• The "it's not X, it's Y" phrasing AI writing leans on is literally one of the patterns being flagged.

This is zero brain cell slop.

If you don't want to spend 2 minutes to write your own opinion, you don't have an opinion worth posting.

The smart stuff

Same AI tools, completely different outcome, because the task fits what AI is actually good at.

Internal prototyping

A prototype's only job is to get tested, then thrown away. Nobody's betting real money on it while it's being tested, so mistakes surface within hours, while they're still cheap to fix.

This is the cleanest use of AI in a business right now: high iteration speed, zero blast radius. It just needs to be fast enough that you can torch 3 bad versions before lunch and still ship the right one by Friday.

Month-end close

Close is painful because it's repetitive and rule-based: match this line item to that invoice, over and over, on a deadline. That's exactly the shape of task AI is good at: the same pattern every month, with a clear right answer each time.

Keep a human reviewing whatever exceptions AI flags. That's the step people skip when they get excited about automation, and it's the one that actually matters. Let AI handle the repetitive matching, and close stops eating a week of your finance team's month.

It’s internal. It’s consistent. It gets reviewed, and the sense checks are well-understood and unlikely to change dramatically in 1, 2, 5 years from now.

Winning customers scalably

Most businesses under $5M lose deals to sloppy process, a lead that sat 4 days, a follow-up that never went out, a proposal that took 2 weeks.

AI is built for exactly that: it never forgets a follow-up and never has a bad Tuesday. Keep humans on the conversations, put AI on the cadence, and the pipeline stops leaking. This is the one we spend our days on.

But use it as an assistant, not a driver.

Customer support triage

Most support tickets are the same 5 questions asked 500 times a week: where's my order, how do I cancel, what's your return policy. That's pattern matching against a knowledge base.

Let AI take the first pass and hand off anything it can't confidently answer. Speed is one win. The bigger one is your best people no longer spending their day on tickets a script could close.

HITL + consistent + scalable.

Sales call notes straight into the CRM

A rep typing notes from memory after a call is flawed by design. They remember what felt important in the moment. Three weeks later, when someone else picks up the deal, that's rarely enough.

AI pulling structured fields straight from the transcript means your CRM reflects what was actually said on the call. No memory required, no gaps to fill in later.

How to actually do it

Three filters, run in order: eliminate, automate, delegate.

Eliminate first. Before you build anything, ask if the task needs to exist at all. A lot of what people automate with AI is work that shouldn't be happening in the first place, like a report no one reads or a status update that could just be a dashboard.

Automate what's left. Once a task survives the eliminate question, hand the repetitive, rule-based chunk of it to AI:

• Bank reconciliation exceptions
• First-pass support replies
• Draft CRM notes from a call transcript

Keep a human reviewing the output until you trust it.

Delegate the whole workflow, later. Only after automation has run clean for a while do you let AI own the full loop unsupervised. Delegation is the reward for a process that's already proven itself, not the starting point.

The exercise: pick one task you did last week purely because it's always been done that way, and run it through eliminate, then automate. If it survives both, build it. If it doesn't survive eliminate, you just saved yourself a coding session.

Before you build anything with AI, ask what breaks when it's wrong.

A prototype breaks an afternoon. A trading bot breaks your savings account. The tools are identical, so the answer to that one question is the entire difference between a build that compounds and one that invoices you later.